Fig. 7: Information extraction pipeline with four main modules.

a Text selection: flash flood episodes selected from the Storm Events Database using keyword filtering and filters on hazard type, time, and location. b LLM prompting: two LLMs used to extract contributing factors from the narratives. c Validation: LLM outputs checked against original narratives to generate the verified results. d Post-processing: extracted information analyzed through word-frequency exploration and normalization.